Fig. 1. Proposed framework for uncertainty-aware GTVp segmentation of OPC patients.

The probabilistic deep learning model ($f\left(\cdot \right)$) with stochastic parameters ($\theta$) distributed according to an approximate posterior distribution ($p$) segments the GTVp, outputs a voxel-level uncertainty map, and quantifies the patient-level uncertainty value ($U$). The patient-level uncertainty is then used to estimate the segmentation quality by checking whether the uncertainty is below or above the predetermined threshold ($\tau$). When the patient-level uncertainty exceeds the threshold, a medical expert will manually inspect and perform corrections to the deep learning segmentation, if necessary. The downstream utilization of the segmentation is then informed by the patient-wise and voxel-wise uncertainties, as well as the patient-wise performance estimate.